Why simple Grok questions trigger a long web search

xAI documents that Grok can run real-time web search and search public X posts so answers can include current information. Free and paid consumer cards on x.ai/pricing both advertise that live search inside their limits. When the search tools fire, you can see a long pause even on a short prompt — that is the tool round-trip, not the model “forgetting” everything.

There is no separate consumer FAQ toggle documented as “never search.”

What is happening

1. On grok.com and the Grok apps, a turn may invoke web search and/or X search when the question looks like it needs current facts.
2. Search covers public pages and public X posts. Private or login-walled pages need a connector you authenticated (Gmail, Drive, and so on).
3. On paid plans, search still draws from the shared weekly usage pool shown under Settings → Usage (product breakdown includes Chat and related spend).

Training cut-off alone does not watch live events. A quiet knowledge question can still trip search if the model chooses the tool for that turn.

What you can do

1. Ask with the facts in the prompt (“using only what I paste below…”) when you want reasoning over your text, not a live crawl.
2. Use DeepSearch only when you want the longer researched pass with sources — see Run DeepSearch.
3. Check Settings → Usage if a burst of searches moved the weekly bar more than you expected.
4. For developer control of when search runs, use the API web search tools — see Add web search to a Grok API request.

Related

• Consumer “does Grok have live internet / live X?” overview: When Grok uses real-time web and X search.
• Weekly pool spend: Spend your Grok weekly usage pool.

Pitfalls

• “Real-time” does not mean every reply streams every public X post — only when search tools run for that turn.
• DeepSearch is a different, longer research flow than a quick web search chip on a normal chat turn.
• Tesla in-car Grok and Grok inside the X apps follow those products’ own surfaces.